DeepMath-7B-M

Model Overview

DeepMath-7B-M is a fine-tuned version of DeepSeek-R1-Distill-Qwen-1.5B on the GSM8K dataset. This model is designed for mathematical reasoning and problem-solving, excelling in arithmetic, algebra, and word problems.

Model Details

• Base Model: DeepSeek-R1-Distill-Qwen-1.5B
• Fine-Tuning Dataset: GSM8K
• Parameters: 1.5 Billion
• Task: Mathematical Question Answering (Math QA)
• Repository: codewithdark/deepmath-7b-m
• Commit Message: "Full merged model for math QA"

Training Details

• Dataset: GSM8K (Grade School Math 8K) - a high-quality dataset for mathematical reasoning
• Fine-Tuning Framework: Hugging Face Transformers & PyTorch
• Optimization Techniques:
  • AdamW Optimizer
  • Learning rate scheduling
  • Gradient accumulation
  • Mixed precision training (FP16)
• Training Steps: Multiple epochs on a high-performance GPU cluster

Capabilities & Performance

DeepMath-7B-M excels in:

• Solving word problems with step-by-step reasoning
• Performing algebraic and arithmetic computations
• Understanding complex problem structures
• Generating structured solutions with explanations

Usage

You can load and use the model via the Hugging Face transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("codewithdark/deepmath-7b-m")
model = AutoModelForCausalLM.from_pretrained("codewithdark/deepmath-7b-m")

input_text = "A farmer has 5 chickens and each lays 3 eggs a day. How many eggs in total after a week?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

• May struggle with extremely complex mathematical proofs
• Performance is limited to the scope of GSM8K-type problems
• Potential biases in training data

Future Work

• Extending training to more diverse math datasets
• Exploring larger models for improved accuracy
• Fine-tuning on physics and higher-level mathematical reasoning datasets

License

This model is released under the mit License.

Citation

If you use this model, please cite:

@misc{DeepMath-7B-M,
  author = {Ahsan},
  title = {DeepMath-7B-M: Fine-Tuned DeepSeek-R1-Distill-Qwen-1.5B on GSM8K},
  year = {2025},
  url = {https://huggingface.co/codewithdark/deepmath-7b-m}
}